How about having the voting machine clearly show the voter's selections before he/she commits to the vote? This could be in the form of a summary page and would give the voter an opportunity to make any necessary changes. Then, once the votes are committed, the voter would get a final verification on a printout. This would avoid all of the complication of having a web-based system like what you proposed.

As for storing the votes on a chip, then transferring the chips manually to the central processor, you might want to add a tamper-resistant coding scheme to make it harder for someone to modify the votes. There are encryption schemes such as RSA that could help reduce the chance for fraud.

I'd want to add some rigid procedures to protect the process even further. For example, before voting starts, I'd want a unique identifier copied to each memory chip in the form of a public encryption key. Every memory chip would be logged in at the beginning of the day and then must be accounted for when the polls close. The idea is to prevent anyone substituting chips once the voting is completed. There might also be the need for some additional verifications such as checksums or CRC to make sure the machine is functioning properly.

Or, we could perhaps go back to paper ballots. Just a thought.